RE: [Bug-gnubg] Wildly different evaluations by CRASHED network

[Øystein O Johansen]

> I've no idea what meta-phi is, but I wonder if some kind of
> fuzzy logic algorithm woud be useful. Take some kind of
> combined output from the RACE and CONTACT networks depending
> on the amount of contact left.

Yes, that's about the same I was thinking. In the input
calculation there's even an input called I_BREAK_CONTACT which
I think can be used for this purpose.

x = I_BREAK_CONTACT

arOutput[i] = f(x) * arRace[i] + ( 1.0 - f(x) ) * arContact[i];

The function f(x) can be found by human trial and error or some
automatic scheme.

The big drawback is: The position must then be evaluated by both
neural nets, which again will slow down the evaluation. This
evaluation must there only be applied in low contact positions
and not anywere else in the game.

-Øystein


Meta pi:
When f(x) is a function of a vector and the result may also be
a vector, the whole thing could be implemented as an other
neural net, would't it? This neural net will control all the
other nets and the net could be trained by supervised training.
This would then be a meta-pi scheme neural net. (I think it's
meta-pi and not meta-phi as I wrote in the first post.)



-------------------------------------------------------------------
The information contained in this message may be CONFIDENTIAL and is
intended for the addressee only. Any unauthorised use, dissemination of the
information or copying of this message is prohibited. If you are not the
addressee, please notify the sender immediately by return e-mail and delete
this message.
Thank you.